Fix Qwen3.5 model config metadata parsing

common/types/model.go

@@ -1,6 +1,7 @@
 package types
 
 import (
+	"encoding/json"
 	"time"
 
 	corev1 "k8s.io/api/core/v1"
@@ -452,6 +453,40 @@ type ModelConfig struct {
 	NumAttentionHeads int      `json:"num_attention_heads"`
 	TorchDtype        string   `json:"torch_dtype"`
 }
+
+type ModelTextConfig struct {
+	NumHiddenLayers   int    `json:"num_hidden_layers"`
+	HiddenSize        int    `json:"hidden_size"`
+	NumAttentionHeads int    `json:"num_attention_heads"`
+	Dtype             string `json:"dtype"`
+	TorchDtype        string `json:"torch_dtype"`
+}
+
+func (c *ModelConfig) UnmarshalJSON(data []byte) error {
+	type modelConfig ModelConfig
+	var cfg struct {
+		modelConfig
+		TextConfig *ModelTextConfig `json:"text_config"`
+	}
+	if err := json.Unmarshal(data, &cfg); err != nil {
+		return err
+	}
+
+	*c = ModelConfig(cfg.modelConfig)
+	if cfg.TextConfig != nil {
+		c.NumHiddenLayers = cfg.TextConfig.NumHiddenLayers
+		c.HiddenSize = cfg.TextConfig.HiddenSize
+		c.NumAttentionHeads = cfg.TextConfig.NumAttentionHeads
+		if cfg.TextConfig.Dtype != "" {
+			c.TorchDtype = cfg.TextConfig.Dtype
+		}
+		if cfg.TextConfig.TorchDtype != "" {
+			c.TorchDtype = cfg.TextConfig.TorchDtype
+		}
+	}
+	return nil
+}
+
 type EngineConfig struct {
 	EngineName      string            `json:"engine_name"`
 	ContainerPort   int               `json:"container_port"`

common/types/model_test.go

@@ -0,0 +1,76 @@
+package types
+
+import (
+	"encoding/json"
+	"testing"
+
+	"github.com/stretchr/testify/require"
+)
+
+func TestModelConfigUnmarshalTopLevelFields(t *testing.T) {
+	content := []byte(`{
+		"architectures": ["Qwen2ForCausalLM"],
+		"attention_dropout": 0.0,
+		"bos_token_id": 151643,
+		"eos_token_id": 151645,
+		"hidden_act": "silu",
+		"hidden_size": 896,
+		"initializer_range": 0.02,
+		"intermediate_size": 4864,
+		"max_position_embeddings": 32768,
+		"max_window_layers": 24,
+		"model_type": "qwen2",
+		"num_attention_heads": 14,
+		"num_hidden_layers": 24,
+		"num_key_value_heads": 2,
+		"rms_norm_eps": 1e-06,
+		"rope_theta": 1000000.0,
+		"sliding_window": 32768,
+		"tie_word_embeddings": true,
+		"torch_dtype": "bfloat16",
+		"transformers_version": "4.40.1",
+		"use_cache": true,
+		"use_sliding_window": false,
+		"vocab_size": 151936
+	}`)
+
+	var config ModelConfig
+	err := json.Unmarshal(content, &config)
+
+	require.NoError(t, err)
+	require.Equal(t, []string{"Qwen2ForCausalLM"}, config.Architectures)
+	require.Equal(t, "qwen2", config.ModelType)
+	require.Equal(t, 24, config.NumHiddenLayers)
+	require.Equal(t, 896, config.HiddenSize)
+	require.Equal(t, 14, config.NumAttentionHeads)
+	require.Equal(t, "bfloat16", config.TorchDtype)
+}
+
+func TestModelConfigUnmarshalTextConfigFields(t *testing.T) {
+	content := []byte(`{
+		"architectures": ["Qwen3_5ForConditionalGeneration"],
+		"model_type": "qwen3_5",
+		"text_config": {
+			"num_hidden_layers": 64,
+			"hidden_size": 5120,
+			"num_attention_heads": 24,
+			"dtype": "bfloat16"
+		},
+		"vision_config": {
+			"depth": 27,
+			"hidden_size": 1152,
+			"num_heads": 16
+		}
+	}`)
+
+	var config ModelConfig
+	err := json.Unmarshal(content, &config)
+
+	require.NoError(t, err)
+	require.Equal(t, []string{"Qwen3_5ForConditionalGeneration"}, config.Architectures)
+	require.Equal(t, "qwen3_5", config.ModelType)
+	require.Equal(t, 64, config.NumHiddenLayers)
+	require.Equal(t, 5120, config.HiddenSize)
+	require.Equal(t, 24, config.NumAttentionHeads)
+	require.Equal(t, "bfloat16", config.TorchDtype)
+}